ホーム / ShortFlingブログ / Writing
この記事は現在、英語版のみ提供しています。英語版を読む →
WritingBest AI Models for Storytelling and Roleplay (Late 2026)
The short answer
If you look at what dedicated storytelling platforms ship rather than what leaderboards say, late 2026 belongs to large open-weight models from two Chinese labs: DeepSeek’s V3.2 and V4 families, and Zhipu’s GLM 4.6 and 4.7, plus fine-tunes built on them. AI Dungeon made a DeepSeek V4 Flash build its default Flash model in August 2026, and NovelAI’s top writing model, Xialong, is a fine-tune of GLM-4.6.
What nobody has is a neutral benchmark for whether a model writes a story you want to keep reading. Coding and reasoning scores are everywhere; fiction quality is mostly judged by feel. So this post does two things: it reports which models storytelling services chose, with dates, and it suggests how to judge them by the only test that matters for fiction, which is the text a reader ends up with. Everything here was checked against official pages as of mid-October 2026, and model lineups change fast.
What the storytelling platforms picked
AI Dungeon’s patch notes are the clearest public record. On August 7, 2026 it added an Optimized Context toggle for DeepSeek 3.2. On August 13 it said an experimental model, DeepSeek V4 Flash 0731, had performed very well in its testing and became a permanent model and the new default for its Flash slot, with the older version still selectable. The same note said GLM 4.6 remains deprecated but available, now on the same context terms as GLM 4.7.
NovelAI went the fine-tuning route. On March 31, 2026 it released Xialong, a fine-tune of GLM-4.6 on its own storytelling dataset. NovelAI’s documentation describes it as its largest fine-tune yet at 355 billion parameters, with a context of 28,672 tokens plus an 8,192-token rollover window. It launched for Opus subscribers only, and NovelAI said it had no immediate plans to bring it to lower tiers.
Neither choice proves a model is the best writer in absolute terms. Platforms weigh cost, speed and how a model behaves under their own prompts. But a dedicated fiction service switching its default is a stronger signal than a general-purpose leaderboard, because it reflects what its own players kept or complained about. For how these tools differ as products, see our guide to NovelAI, Sudowrite and AI Dungeon and our explainer on Latitude’s new Voyage platform.
The models, briefly
DeepSeek V4. DeepSeek released V4 as an open-weight preview on April 24, 2026, in two sizes: V4-Pro with 1.6 trillion total and 49 billion active parameters, and V4-Flash with 284 billion total and 13 billion active, both with a one-million-token context. On September 10, 2026 DeepSeek released V4.1-Flash and retired V4-Flash on its own API, routing old requests to the new model, and from September 14 it routed V4-Pro requests to V4.1-Flash too. That is why names get confusing: a platform hosting its own copy of an open-weight build, like AI Dungeon’s dated V4 Flash 0731, can keep serving a version DeepSeek’s own API no longer offers.
GLM 4.7. Zhipu’s developer documentation markets GLM-4.7 mainly on coding and agent work, but it also claims more vivid, descriptive prose and consistent adherence to world-building and character archetypes in role-play, and lists a 200K context. The same documentation now features newer GLM-5 models, so 4.6 and 4.7 are no longer Zhipu’s latest; on story platforms they remain in use.
Xialong. NovelAI’s pitch for its GLM-4.6 fine-tune is specifically about fiction behavior rather than benchmarks. It says Xialong was trained not to repeat itself or fall into loops, and that, unlike models whose probabilities collapse after reinforcement learning, it gives genuinely different outcomes on retries and more varied character names. Those are NovelAI’s claims, but they name exactly the problems fiction writers complain about most.
How to judge a model by what the reader reads
Repetition is the first test. A common complaint about AI fiction is loops: the same gesture, the same phrase, the same paragraph shape every few turns. Generate a long scene and read it start to finish, not turn by turn. If you notice the tics, a reader will too.
Retry variety is the second. Ask for the same beat five times. If you get five near-identical answers, the model will be a weak brainstorming partner, however polished each one looks. Third, check continuity across length: whether a name, scar or promise from chapter two survives to chapter ten. Large context windows help, but they are not the same as paying attention, and drift over long sessions is a common complaint about many models.
Finally, judge voice. A model that defaults to the same polite, summarizing register for every character will flatten a romance or a thriller however capable it is on paper. Our post on whether AI can write good romance goes deeper on that point.
Local, cloud or an app?
Open weights do not mean a model runs on your laptop. The models above range from hundreds of billions of parameters to well over a trillion, which puts the full versions out of reach of a typical home computer. In practice most people use them through a hosted API or through an app that runs them for you.
Cloud APIs are cheap per word, though prices move. As of mid-October 2026, DeepSeek’s pricing page listed its Flash model at $1.20 per million output tokens at peak times, with off-peak rates at half that. The trade-off is setup: you need an account, an API key and a front end, and you are responsible for prompts, context management and each provider’s usage policies. Our comparison of Janitor AI, SillyTavern and NovelAI covers that do-it-yourself route. An app does the setup for you but chooses the model.
Where ShortFling fits
ShortFling is a reading and writing platform, not an AI chat or role-play app. If you want help writing a story rather than running a campaign, the writing studio has an AI editor that is free within a daily allowance, with no API keys or front ends to configure. You still write the story: work must be your own original writing with adult characters only, and pressing Publish submits it for review, so it goes public only after approval.
And if you want to judge AI-assisted prose by reading it, the fiction library has finished serials in separate SFW and 18+ NSFW sections, with NSFW availability depending on your region and platform.
FAQ
What is the best AI model for roleplay in late 2026?
There is no neutral fiction benchmark. Going by what storytelling platforms ship, the main choices are DeepSeek’s V3.2 and V4 families and GLM 4.6 and 4.7, plus fine-tunes such as NovelAI’s Xialong. AI Dungeon made DeepSeek V4 Flash 0731 its default Flash model on August 13, 2026.
What is NovelAI’s Xialong?
Xialong is NovelAI’s writing model released on March 31, 2026: a 355-billion-parameter fine-tune of GLM-4.6 on NovelAI’s storytelling data, trained to reduce repetition. It launched for Opus subscribers only.
Is DeepSeek V4 Flash still available?
On DeepSeek’s own API, V4-Flash was retired on September 10, 2026 and requests now go to V4.1-Flash. Platforms that host their own copy of an open-weight build, such as AI Dungeon’s V4 Flash 0731, can keep offering it.
Can I run these models locally?
The full versions are very large, from hundreds of billions of parameters upward, so they are out of reach of typical home hardware. Most people use them through a hosted API or an app.
Sources
- AI Dungeon: What’s New (DeepSeek V4 Flash 0731 made default Flash on Aug 13, 2026; GLM 4.6/4.7 notes; DeepSeek 3.2 toggle on Aug 7, 2026) (checked 2026-10)
- NovelAI blog: Welcome NovelAI’s newest writing model, Xialong (Mar 31, 2026; GLM-4.6 fine-tune, Opus only, anti-repetition training) (checked 2026-10)
- NovelAI documentation: Text models (Xialong parameters and context size) (checked 2026-10)
- DeepSeek API docs: DeepSeek V4 Preview release (Apr 24, 2026; model sizes, 1M context, open weights) (checked 2026-10)
- DeepSeek API docs: DeepSeek-V4.1-Flash release (Sept 10, 2026; V4-Flash retired, V4-Pro routing) (checked 2026-10)
- DeepSeek API docs: Models and pricing (Flash output price, off-peak rates) (checked 2026-10)
- Z.ai developer docs: GLM-4.7 overview (writing and role-play claims, 200K context) (checked 2026-10)




